Nvidia booked $96.2B in its fiscal second quarter, the company reported after the U.S. close on August 26. The same quarter a year earlier came in at $46.7B. Analysts had modeled $92.2B; Nvidia’s own guidance was $91B plus or minus 2%. Both were left behind.
Then Huang wrote the sentence that matters. “AI has reached its inflection point. It’s doing useful work. Its tokens are productive and profitable. Now, compute is revenue,” he said in the release.
That is not a flourish. It is a claim about substitution, and it is the seller making it. For two years the companies writing “AI-driven efficiency” into layoff filings have all been buyers. The seller talked about demand, capacity, and lead times. This quarter the seller skipped to the conclusion: your revenue should track compute, not the other thing it used to track.
The line Nvidia split in two
Data center revenue came in at $89B against a $85.7B estimate. Last quarter the segment produced $75.2B. A year ago, $39.1B.
The mix is the news. For the first time, Nvidia broke data center revenue into two disclosed lines: $48.7B from hyperscalers, and $40.3B from what the company now calls ACIE — AI clouds, industrial, and enterprise. Until this quarter those buyers sat inside one aggregate number and outsiders guessed at the ratio. Splitting it out tells the market something specific: close to half the compute Nvidia sold this quarter did not go to the five large cloud platforms.
The rest of the print runs the same direction. Third-quarter guidance is $108B plus or minus 2%, against a $104.2B consensus, per CoinDesk. Huang told analysts to expect roughly 70% revenue growth in fiscal 2028, well above Wall Street. Non-GAAP EPS landed at $2.22 against a $2.06 to $2.09 range, with one footnote worth carrying: Nvidia began including stock-based compensation in its non-GAAP figures this quarter, so it is not the same ruler used in prior years.
Vera Rubin is ramping into full production and starts contributing in the fiscal third quarter. Amazon Web Services will buy 2 million GPUs and adopt Nvidia’s new Vera CPU alongside them. That last clause carries more than it looks like. The CPU is the slot Nvidia never owned in a rack it otherwise designed end to end. Vera shipping with Rubin means the company now books the whole rack.
Whose P&L pays the $89B
Eighty-nine billion dollars of compute in one quarter comes off somebody else’s income statement. We wrote that side of the trade last week: Alibaba’s capital expenditure rose 75% year over year while net profit fell 76% (Alibaba’s capex rose 75%. Its profit fell 76%.). A buyer’s books look like that so a seller’s books can look like $96.2B. Same transaction, two ends.
The $40.3B ACIE line is the genuinely new information. Hyperscaler compute demand has been modeled to death for two years and is priced in. Enterprises and sovereign buyers purchasing compute directly is a different fact, because it means the substitution has walked out of the technology sector and into ordinary corporate IT budgets. Companies break out a line when it gets too large to bury, and when they want the market to value it at a higher durability multiple. Both apply here.
For a cohort point, look at physical automation, where the same money is moving the same way. A3’s North American robot order data for the first half, published August 11, showed units up 2.0% and order value up 6.6% — buyers spending three times faster than they were adding machines. Data center or factory floor, the budget is concentrating in the compute and software layer rather than in more physical units. The same dollar buys “smarter,” not “more.”
What “compute is revenue” means for people
If a company’s revenue scales with compute, then on the budget sheet the next hire competes with the next block of GPU-hours for the same money. CFOs have been circling this for two years with “operating leverage” and “structural efficiency.” Huang put it in three words in an earnings release.
The number to watch is the $40.3B, not the $96.2B. The headline total says cloud platforms are building capacity, which maps to data center construction and operations work: electricians, HVAC, substation crews, facilities techs, civil trades. Those roles have been net additive for two years and this guide extends the run. ACIE says non-technology enterprises are now buying compute on their own account, which maps to jobs inside those enterprises: IT operations, internal platform teams, and the analyst layer whose output can be repriced as a token cost. Oracle’s second round of cuts in August sits squarely in that category (Oracle Plans a Second Round of Cuts With Its Budget Nearly Spent).
There is a usable test in that split. Ask whether your role builds compute or consumes it. On the build side, demand tracks the $108B guide upward. On the consume side, demand tracks cost-per-task downward. One earnings release, pointing two directions at once, depending on which side of the rack you sit.
The timing is derivable. The fiscal third quarter is when Rubin ramps, so the $108B guide converts into installed capacity across this quarter and the next. Enterprise cost-per-task follows that curve down. Headcount decisions follow the budget cycle after that, typically one to two quarters later. Which puts the first felt effect of tonight’s numbers on ordinary corporate staff somewhere in the first half of fiscal 2028. The gap is not a cushion. It is the preparation window.
One more detail belongs next to this. Nvidia is not the company doing the layoffs, but its own hiring has changed shape: last week it paid $6B for a non-exclusive Poolside licence and extended offers to 109 people (Nvidia paid $6B for a licence and made offers to 109 people). An entire technical core moved without an acquisition and without a hiring ramp. The company selling the compute is already running the operating model that the compute makes possible. Its customers are learning it next.
